Mo' Data Mo' Problems: How Data Composition Compromises Scaling Properties

Judy Hanwen Shen, Inioluwa Deborah Raji, Irene Y. Chen

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: data composition, tabular data, measuring data, dataset bias, disparities

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Why more data may not improve performance or bias

Abstract: The accumulation of data in the machine learning setting is often presented as a panacea to address its many modeling problems---including issues with correctness, robustness, and bias. But when does adding more data help, and when does it hinder progress on desired model outcomes? We model data accumulation from multiple sources and present analysis of two practical strategies that result the addition of more data degrading overall model performance. We then demonstrate empirically on three real-world datasets that adding training data can result in reduced overall accuracy and reduced worst-subgroup performance while introducing further accuracy disparities between subgroups. We use a simple heuristic for determining when the accumulation of more data may worsen the issues the additional data is meant to solve. We conclude with a discussion on considerations for data collection and suggestions for studying data composition in the age of increasingly large models.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7727